US 12,482,004 B2
Carbon footprint estimation using foundation model
Jagabondhu Hazra, Bangalore (IN); Manikandan Padmanaban, Chennai (IN); Ayush Jain, Lucknow (IN); Ranjini Bangalore Guruprasad, Bangalore (IN); and Shantanu R. Godbole, Bangalore (IN)
Assigned to International Business Machines Corporation, Armonk, NY (US)
Filed by International Business Machines Corporation, Armonk, NY (US)
Filed on Sep. 25, 2023, as Appl. No. 18/473,435.
Prior Publication US 2025/0104093 A1, Mar. 27, 2025
Int. Cl. G06Q 30/018 (2023.01)
CPC G06Q 30/018 (2013.01) 20 Claims
OG exemplary drawing
 
1. A method of estimating Scope 3 emissions, the method comprising:
generating embeddings based on enterprise financial transaction data, enterprise metadata, and crowd source data, wherein generating the embeddings comprises:
performing a tokenization operation on training data that has units of a first size and that represent the enterprise financial transaction data, the enterprise metadata, and the crowd source data, the tokenization operation dividing the training data into tokens having a second size that is less than the first size to generate the embeddings;
generating sector wise carbon-aware spatio-temporal weights indicative of an estimated level of Scope 3 emissions produced by a plurality of corresponding commodity sectors;
passing the embeddings and sector wise carbon-aware spatio-temporal weights through a pre-trained carbon-aware foundation model (FM) to generate predicted emissions data;
inputting the predicted emissions data into a carbon-aware loss function to compute a loss-function value (Le), the carbon-aware loss function defined using the carbon-aware weights for sectors, geography, and time period of transaction data, wherein the transaction data includes the enterprise financial transaction data, the enterprise metadata, and the crowd source data;
repeatedly training the pre-trained carbon-aware FM until the loss-function value (Lθ) is less than or equal to a target threshold (Th);
generating a carbon-aware natural language processing (NLP) foundation model (FM) that is trained according to the embeddings and the sector wise carbon-aware spatio-temporal weights in response to the loss-function value (Lθ) being less than or equal to a target threshold (Th);
inputting into the NLP FM user-generated data indicating at least one target commodity sector and spend data associated with the target commodity sector; and
outputting from the NLP FM an estimation of the Scope 3 emissions based on the least one target commodity sector and the spend data.